gender gap in wage returns to job tenure and experience · of gender differences in wage returns to...

r.com/locate/econbase
Labour Economics 15 (2008) 1296–1316
www.elsevie

Gender gap in wage returns to job tenure and experience☆

Lalith Munasinghe b,⁎, Tania Reif a, Alice Henriques b

a Citigroup, 388 Greenwich Steet, New York, NY 10013b Department of Economics, Barnard College, Columbia University, 3009 Broadway, New York, NY 10027

Received 26 July 2004; received in revised form 15 December 2007; accepted 19 December 2007Available online 28 December 2007

Abstact

We present empirical evidence on gender disparities in wage returns to job tenure and experience. Wefind that the overall wage return on an extra year of labor market experience is lower for women than men.A decomposition analysis shows that the wage return to job tenure is substantially lower for women thanmen, and that the wage return to experience is higher for women than men. These gender differences arerobust to various estimation procedures and are especially pronounced for more educated workers. Wehypothesize that these observed gender disparities in wage returns are driven by the fact that women are lessattached to their jobs than men. We present some supportive evidence for our hypothesis, namely, thatwomen are more likely to quit their jobs, receive substantially fewer hours of company provided training,and a much higher fraction of women expect to not work at age 35 due to family related reasons.© 2008 Elsevier B.V. All rights reserved.

JEL Classification: J16; J24; J31; J41Keyword: Gender Wage Gap; Returns to Tenure; Returns to Experience

1. Introduction

The changing role of women in the labor market is an unparalleled transformation of thetwentieth century. The dramatic increase in labor force participation, catch up in educationalattainment, legal battles of equal pay for equal work, and the narrowing, but still persistent, gender

☆ We thank Satyajit Bose, Joyce Jacobsen, Audrey Light, Jacob Mincer, Brendan O'Flaherty, Steffen Reichold, RenaRosenberg, Rajiv Sethi, and Nachum Sicherman for many helpful comments. We also thank two anonymous referees andthe editor for excellent comments on previous versions of the paper.⁎ Corresponding author. Tel.: +1 212 854 5652.E-mail addresses: [email protected] (L. Munasinghe), [email protected] (T. Reif),

[email protected] (A. Henriques).

0927-5371/$ - see front matter © 2008 Elsevier B.V. All rights reserved.doi:10.1016/j.labeco.2007.12.003

mailto:[email protected]



http://dx.doi.org/10.1016/j.labeco.2007.12.003

1297L. Munasinghe et al. / Labour Economics 15 (2008) 1296–1316

wage gap remain key research issues across various disciplines. A major focus in the economicsliterature is the earnings disparity across women and men. Recent empirical studies (Blau andKahn, 1997; O'Neill and Polachek, 1993; Wellington, 1993) attribute an important part of thenarrowing of the gender wage gap to themarked increase in both women's labor force participationrates and accumulation of labor market experience. In this paper we present an empirical analysisof gender differences in wage returns to an extra year of job tenure and labor market experience.The decomposition of wage returns into a firm-specific and general component allows us toidentify a potentially important source of gender disparities in earnings dynamics.

In our empirical analysis we use the National Longitudinal Surveys of Youth (NLSY) toestimate wage returns to tenure and experience. Our sample consists of a panel of young womenand men that spans the first decade and a half of their careers. We find that the sum of wage returnsto tenure and general experience is lower for women than men. The decomposition analysis,however, shows that the wage return to tenure is substantially lower for women than men, and thatthe wage return to general experience is slightly higher for women than men. These findings arerobust to various estimation procedures and model specifications, and are especially strongamong more educated workers.

Despite their growing attachment to the labor market, women are likely to be less attached totheir employers and jobs compared to men especially during the early part of their careers.Women in their twenties and thirties experience life cycle events such as marriage, childbirth, andfamily care responsibilities that make them more prone to employment interruptions and gaps. Ifsuch life cycle events lead to less durable employment relationships for women then a variety oftheoretical considerations imply lower wage returns to tenure and higher wage returns toexperience for women compared to their male counterparts. First, if women are more likely to quittheir jobs then the theory of firm-specific human capital investments predicts that women aremore likely to invest in general skills that are portable than in firm-specific skills. Such strategicconsiderations in investments in human capital can explain why returns to tenure are lower andreturns to experience are higher for women than men. Second, bonding and selection models(Lazear, 1981; Salop and Salop, 1976) that imply back-loaded compensation designs are alsoconsistent with our findings. An important corollary of these models is that women who anticipatemore frequent job changes in the future will have less incentives to select into jobs with back-loaded compensation.1 Third, lower returns to tenure for women could be due to employerdiscrimination based on rational expectations about women's attachment to jobs. For example, ifemployers engage in statistical discrimination then the perception of women as “less” stableworkers could lead employers to systematically not hire women into jobs with opportunities ofspecific training or learning.2 These theoretical considerations suggest that if women are indeedless attached to their employers then wage returns to tenure are likely to be lower, and wagereturns to experience possibly higher for women than men.

Since gender disparities in returns to tenure and experience can be explained by the hypothesisthat women have weaker ties to their employers, we begin our empirical section with an analysisof gender disparities in labor market and job attachments. First, we present a turnover analysis thatshows women are more likely to quit their jobs than their male counterparts. Second, we use a

1 Note that Hersch and Reagan (1997), citing evidence that returns to tenure tend to be greater for female than for maleworkers, develop an agency model of wage contracts that shows efficient wage-tenure profiles are steeper for womenthan men as a direct result of their shorter working life.2 We do not provide a summary of the extensive literature on gender discrimination here. For an excellent survey of the

literature see the article by Altonji and Blank in the Handbook of Labor Economics (1999).

1298 L. Munasinghe et al. / Labour Economics 15 (2008) 1296–1316

small sample on formal training programs and show that women receive substantially fewer hoursof company provided training compared to men. Third, we explore a survey instrument onexpectations about future labor force participation, and find that a nontrivial fraction of women intheir late teens and early twenties expect not to be in the labor force at age 35 due to family relatedreasons. By comparison hardly any men expect to be out of the labor force for family relatedreasons at age 35. All of this evidence points to the fact that women expect and are indeed morelikely to be less attached to their employers.

The implications of our hypothesis about weaker ties to employers among women ispredicated on various strategic human capital investment considerations. As a consequence, wepresent all of our analyses on gender disparities for high and low educated workers separately. Ifeducation is a proxy for lower discount rates and greater trainability then we would expect genderdifferentials to be more pronounced for our more highly educated sample. The observed genderdisparities in wage returns, and in labor force and job attachments are consistently stronger amongour sample of more educated workers.

The remainder of the paper is organized as follows. In the next section, entitled “RelatedLiterature,” we discuss a variety of findings on the gender wage gap. In Section 3 we present ourempirical results in various sub sections. First, we describe the NLSY data and present summarystatistics of the key variables. Second, we present preliminary evidence on gender differences inquit rates, training incidence and duration, and on expectations related to future labor forceparticipation. Third, we outline our estimation framework. Fourth, we present and discuss ourestimates of within-job wage growth and returns to tenure and experience across women and men.Finally, we address some potential biases in the estimation of gender wage returns. Section 4concludes with a brief summary and discussion of further extensions.

2. Related literature

Prior studies have documented the effects of job tenure and experience on the gender wagegap, where the primary focus has been to assess the contribution of gender differences in actualexperience and labor force interruptions (Light and Ureta, 1990; Kim and Polachek, 1994;Wellington, 1993; Filer, 1993). These studies conclude that gender differences in participationrates, timing of work experience, and accumulated experience are not only key determinants ofearnings, but the convergence of these work history characteristics across women and men is animportant source of the recent narrowing of the gender wage gap. In this section we briefly reviewthis literature and highlight some findings that suggest wage returns to tenure are higher forwomen than men, contrary to what we find in our paper.

Light and Ureta (1995) find that the overall wage returns to tenure and experience are higher formen than women and that women experience a smaller drop in wages when they re-enter the labormarket after a nonworking spell compared to their male counterparts. Somewhat surprisingly theyalso find that the wage returns to tenure are almost three times larger for women compared tomen. Itshould be noted however that the magnitudes of their estimates of returns to tenure are extremelysmall for both men and women. These low estimates could be due to the instrumental variables (IV)model they implement, since IV typically over estimates the returns to experience and underestimates the returns to tenure.3 Moreover, the authors use the National Longitudinal Survey wheremissing data on weeks worked can cause measurement error of the tenure variable (Spivey, 1995),

3 See Topel (1991) for a careful discussion of the reasons why returns on tenure are consistently under estimated withIV. Also see Altonji and Williams (2005) for a rebuttal.


and their sample is restricted to very young workers aged 24 to 30 years.We try to attenuate some ofthese restrictions in our analysis by implementing multiple estimation methods, and by using theNLSY where measurement error is less likely to occur and the oldest worker is 46 years of age.

Becker and Lindsay (1994) also find that wage-tenure profiles are steeper for women than menusing the Panel Study of Income Dynamics from 1968 to 1987. Their analysis is based on a samplerestricted to workers who stay with the same employer for at least five years.4 This selectioncriterion of course over samples workers who are more attached to their jobs, and it is likely thatamong women this criterion would select the most durable employment relations. Hence it is notentirely surprising if wage returns to tenure are comparable or higher for women compared to menfor a sample with this selection criterion. In our paper we present evidence on gender disparities inquit rates to assess whether women are in fact less attached to their employers.

In a recent study Banzhaf (2005) uses the same NLSY data and implements a structural modelto address whether job turnover and wage growth vary by education and gender. She finds that inthe first few years of tenure, women experience a higher return to tenure among a sample of moreeducated workers. However, after three years of tenure men continue to experience increasingwages with tenure while women experience declining wages.

Gender disparities in wage returns to tenure also have a direct link to the recent literature on the“family gap” in earnings for women with children.5 Differential returns to marital and parentalstatus explain a large fraction (in the order of 40%–50%) of the gender wage gap (Waldfogel,1998a). One explanation is that employers may discriminate against women with children interms of company provided training and promotions. A second noteworthy reason could be thelack of access to job-protected maternity leave that impedes progress of mothers in labor marketsthat value experience, tenure, and a good match between employer and employee (Waldfogel,1998a,b). Lower wage returns to tenure for women are entirely consistent with this family penaltyliterature. An important policy issue, stressed in this literature, is the ramification of legislationsuch as the Family and Medical Leave Act (FMLA) on the family penalty.6 A potentiallyinteresting question for future research would be to ask whether the returns to tenure for womenincrease in the post 1993 period after the enactment of FMLA.

3. Empirical application

3.1. NLSY data, sample restrictions, and summary statistics

We use panel data from the NLSY (1979–1994) to estimate the returns to experience andtenure across women and men. Starting in 1979, 12,686 male and female youths, aged 14 to 22,

4 Becker and Lindsay (1994) present an extension of Hashimoto's (1981) model of financing firm-specific investmentsto show that women will bear a higher cost of these investments because they are more likely to turn over. As aconsequence, they argue that women will have steeper wage-tenure profiles than men. This result is sensitive to theimplicit assumption that all workers receive the same amount of specific investment. In a less restrictive theoreticalframework where the investments are endogenous, this result is unlikely to hold. As we stress in this paper, if women aremore likely to separate from their employers then they are less likely to invest in specific skills and in jobs withbackloaded compensation.5 See Waldfogel (1998b) for an excellent review of this literature.6 Waldfogel (1998a) presents evidence showing that women who are allowed maternity leave are more likely to return

to their previous employers. Clearly such mandates will increase expected job duration and thus increase the likelihood ofwomen receiving job specific investment opportunities. These effects will not only pertain to mothers but more generallyto women since expectations about motherhood especially from the employers perspective is the more likely determinantof job specific investment opportunities.


were interviewed annually until 1994, the last year of our panel.7 Although the NLSY recordsinformation about multiple jobs, we only consider the CPS designated job both because it is themain (or more recent) job, and because more information is available about the CPS job. Thewage measure we use is the hourly wage deflated by the consumer price index (from theEconomic Report of the President) with 1987 as the base year.

Our sample is restricted to white males and females who are neither self-employed noremployed in the agricultural or government sector. After excluding observations with missingvalues for our analysis variables we have 34,524 person-year observations. We apply othersample restrictions to exclude erroneous wage information and other data inconsistencies. First,we drop 764 person-year observations due to the following criteria: (a) if reported hourly wagesare in excess of $100 or less than $1.00, (b) if real wage losses are in excess of 30% or real wagegains are in excess of 300% across contiguous survey years when the worker is on the same job,and (c) if nominal wage increases are in excess of 200% and is immediately followed by nominalwage decreases in excess of 50%, and vice versa.8 Excluding these observations raise ourestimates of wage returns to tenure substantially and decrease our estimates of wage returns toexperience marginally. For example, consider the 5-year cumulative returns for those with highschool or less years of schooling using one of our estimation procedures (Topel-IV): when weexclude these observations the 5-year cumulative returns to tenure increase from .026 to .176 forwomen and from .130 to .241 for men; and the 5-year cumulative returns to experience decreasefrom .230 to .170 for women and from .197 to .160 for men.9 Clearly, excluding theseobservations work against our main conclusion that women have lower returns to tenure andhigher returns to experience compared to their male counterparts. But we believe that these wagedata are erroneous and should be excluded from our sample.

We make two further sample restrictions that make a minimal impact on our estimates. Weexclude all individuals with less than 9 or more than 16 completed years of schooling. Althoughwe lose 2,078 person-year observations, our estimates of returns to tenure and experience hardlychange at all. finally we exclude a further 317 person-year observations due to inconsistenciesbetween our tenure and experience variables. Note the construction of the experience variable isbased on actual weeks worked and the cumulative tenure variable is based on the start andstopping dates of work with a specific employer. Due to employment breaks, it is possible that thetenure variable overstates the actual weeks worked. We implement a simple correction bydropping person-year observations where the changes in tenure exceed the changes in experiencefrom one survey period to the next. Almost all of the coefficient estimates remain exactly the sameafter this exclusion.10 Our final data sample consists of 31,365 person-year observations.

Table 1 shows the means and standard deviations (in parentheses) for the key variables bygender and education groups. The low education group is restricted to those between 9 and

7 Since 1994 the NLSY data are collected only every two years, and as a result some inconsistencies arise in annualearnings in the off years. To avoid such problems we restrict our analysis to the years 1979–1994 of the NLSY.8 Note that the NLSY does not edit the values of hourly wages which are constructed from reported information about

usual wages, time period of pay, and usual hours worked over that time period. Extreme recorded values for hourly wagestop $25,000 and can be less than $0.01. One noteworthy point is that excluding observations with negative nominalwithin-job wage growth increases our estimates of within-job wage growth to levels comparable to Topel's (1991)estimates.9 The parallel set of estimates for the more than high school group are: 5-year cumulative returns to tenure increase

from − .040 to .086 for women and from .155 to .258 for men; and 5-year cumulative returns to experience decrease from.453 to .421 for women and from .294 to .311 for men.10 We thank Audrey Light for alerting us to this issue.

Table 1Key Variables and Descriptive Statistics, National Longitudinal Surveys of Youth (NLSY) 1979–1994

Variable High School or Less More than High School

Male Female Male Female

Real Hourly Wages (1987 Dollars) 7.91 (3.44) 5.94 (2.51) 10.49 (5.26) 8.34 (4.18)Age (Years) 25.71 (4.05) 25.49 (4.17) 27.47 (3.52) 27.04 (3.56)Education (Completed Years) 11.47 (0.94) 11.60 (0.84) 14.75 (1.26) 14.73 (1.23)Tenure (Years) 2.73 (3.10) 2.39 (2.75) 2.73 (2.79) 2.51 (2.73)Actual Experience (Years) 6.66 (3.97) 5.90 (3.71) 7.78 (3.63) 7.32 (3.58)Married, Spouse Present 0.45 (0.50) 0.49 (0.50) 0.45 (0.50) 0.47 (0.50)Union (Wages Set by Collective Bargaining) 0.21 (0.41) 0.12 (0.32) 0.10 (0.30) 0.07 (0.26)Usual Weekly Hours Worked 42.80 (8.97) 36.73 (8.89) 43.97 (9.15) 38.36 (9.42)Health Limits Amount or Type of Work 0.02 (0.12) 0.03 (0.16) 0.01 (0.11) 0.02 (0.15)AFQT Score 42.52 (23.81) 42.09 (21.80) 70.97 (21.98) 67.16 (20.59)Number of Individuals Number of Observations 1,892 11,245 1,807 9,372 1,040 5,064 1,201 5.645


12 years of completed schooling, and it includes high school graduates. The high education groupis restricted to those between 13 and 16 years of completed schooling, including collegegraduates.11 The mean ages are about the same across women and men within each of these twoeducation groups. The most striking difference is in the mean hourly wages across women andmen. Among the more educated workers, the mean wage for men is about $11 and for women it isabout $8.50. That translates into an almost 30% gross wage premium for men. Among the lesseducated workers, this premium is almost 35%. This gender wage discrepancy is particularlystriking given that the means of some standard human capital measures such as completed yearsof schooling and AFQT scores are very similar across women and men within each educationgroup. Men also have slightly higher mean tenure and mean experience. Within education groups,women work in jobs that are less likely to be union represented and they work fewer hours perweek.

In the next section we turn to our econometric framework to estimate wage returns to tenureand experience.

3.2. Estimation framework

In Section 3.4 below we present estimates of wage returns to tenure and experience from threedifferent models, each of which have been extensively used and discussed in the literature. Thissection briefly describes the estimation strategies underlying each of these models and discusstheir comparative advantages and disadvantages. Before we get to the specifics of these models,consider the standard human capital model of wage determination which underlies all three ofthese models:

yijt ¼ Xijtb1 þ Tijtb2 þ Fijtgþ eijt; ð1Þ

11 Our wage analyses are presented for these two education groups separately. This classification follows Royalty (1998)who presents her mobility analyses along the same classification of low and high education groups. Alternativespecifications yield qualitatively similar results.


where yijt is the log wage for individual i on job j at time t, X is total labor market experience, T istenure (current job seniority) and Fijt is a vector of other determinants of wages including higherorder terms of experience and tenure. Henceforth we will drop Fijtγ for expository convenience.Parameters β1 and β2 represent the returns to an additional year of experience and tenure,respectively. The common interpretation of these coefficients is that β1 represents the return ongeneral human capital and β2 represents the return on accumulated job-specific capital.

The error term is decomposed as:

eijt ¼ Ai þ /ijþ gijt þ vit;

where μi is a fixed individual specific error component, ϕij is a fixed job match specific errorcomponent,ηijt is a time varying jobmatch specific component, and vit is noise. Various hypotheticalcorrelations between these error components and the two key independent variables — tenure andexperience— give rise to bias estimates of returns to tenure and experience using standard OLS.12

These correlations are systematic because they are the outcome of optimizing search behavior, i.e.,workers searching for, finding and maintaining good (high wage) employment matches, and onaccount of individual heterogeneity. The OLS estimates are further confounded because the bias onthe tenure and experience coefficients cannot be signed.13

To formalize these potential biases consider the following decomposition of the unobservables:

eijt ¼ /ijt þ vit; ð2Þ

where ϕijt represents the stochastic component of wages that is idiosyncratic to a specificemployment relationship, and a high value of ϕ implies a “good match.” Now consider thecorrelation of experience and tenure with this unobservable component:

/ijt ¼ Xijtb1 þ Tijtb2 þ uijt ð3Þ

The job search process and increasing within-job wages imply that b1N0; but the sign of b2 isambiguous. However, as Topel (1991) claims b1+b2N0 since good jobs survive. Thesetheoretical correlations are the source of potential biases that the various estimators presentedbelow try to circumvent.

We consider three econometric models — due to Altonji and Shakotko (1987) and Topel(1991) — that are designed to address these potential biases resulting from job match hetero-geneity and individual heterogeneity. The objective of all of these models is to provide tenureand experience coefficients that are more reflective of the “true” wage returns to tenure and

12 The controversy about estimating returns on tenure has much to do with the current consensus among empirical laboreconomists that the estimated tenure effect is either small or absent (Altonji and Williams, 2005), while the negativeeffect of tenure on job turnover remains one of the most robust and widely documented findings in empirical labor. Sincethe size of the tenure effect on wages is typically interpreted as an indicator of the importance of firm-specific skills or ofback-loaded compensation schemes, the small estimated coefficient appears to cast doubt on the workhorse theories ofcompensation and turnover. However, it is important to keep in mind that in the presence of firm-specific rents, Becker-type sharing models and dynamically consistent wage policies (Munasinghe, 2006) imply flatter wage-tenure profilescompared to the underlying productivity-tenure profiles. Hence small tenure effects on wages are consistent with highlevels of firm-specific skill accumulation. These considerations suggest that even if the estimated gender differentials inreturns on tenure are small, it may nevertheless represent much larger differences in firm-specific skill accumulationacross women and men.13 The nature of these baises are explained in much greater detailed in Topet (1991), and Altonji and Williams (2005).


experience. Since all of these models have been extensively used and discussed in the literature,we will only briefly describe the estimation strategies underlying each of them. Although the sizeof the point estimates differ depending on which estimator is implemented (see Altonji andWilliams (2005) for a more recent comparative assessment of these methods), in Section 3.4below we present estimates from all three procedures and show that the gender disparitiesin returns to tenure and experience remain robust especially for our sample of more educatedworkers.

The first model we consider is Topel's (1991) two-step estimator (Topel-2S, from now on),which provides an unbiased estimate of the overall returns to an extra year of labor marketexperience, a lower bound of the returns to tenure, and an upper bound of the returns to generallabor market experience. Although the second step estimates of returns to experience and tenureare themselves not unbiased, the direction of bias is implied by theory and the bias is symmetricalacross the two estimates. An important issue that we explore in Section 3.5 below is whether thisbias is different across women and men.

The first step of Topel-2S is to estimate within-job wage growth from the first difference ofEq. (1) for individuals who do not change jobs, which eliminates fixed job and individual effects:

yijt � yijt�1 ¼ b1 þ b2 þ eijt � eijt�1 ð4ÞIf eijt − eijt−1 has zero mean and is serially independent then OLS applied to Eq. (4) above will

yield a consistent and efficient estimate of within-job wage growth, β1+β2.14

The second step entails the estimation of the returns to experience and tenure. Since X=X0+Twe can rewrite Eq. (1) as follows:

y� T bb1 þ b2� �

¼ X0b1 þ �; ð5Þ

where �=e+T ((β1+β2)−T bb1 þ b2� �

), and bb1 þ b2 is the estimate of the sum of the returns onexperience and tenure from the first step. Eqs. (4) and (5) define Topel's two-step model ofestimating returns to tenure and experience. Provided b1N0 and (b1+b2)N0 for reasonsmentioned previously, we estimate an upper bound for experience and a lower bound for tenure.Since these are well settled properties of job search models, Topel's two-step method leads tobiased estimates where the direction of bias is well defined.

Note that the OLS estimatebb1 from Eq. (5) above is an overestimate of the true returns toexperience because X0 is correlated with �. Hence our estimate of returns on tenure,bb2 ≡bb1 þ b2−bb1 , is a lower bound of the average return on seniority. These implications of thetwo-step procedure can be explicitly derived by applying OLS to (4) and (5) and computingthe expectations ofbb1 andbb2 as follows:

E bb1� �¼ b1 þ b1 þ gX0T b1 þ b2ð Þ ð6Þ

E bb2� �¼ b2 � b1 � gX0T b1 þ b2ð Þ; ð7Þ

where γ X0 T is the least squares coefficient from a regression of tenure on initial experience,X0. Notice that the biases of the estimates of the returns to tenure and experience areequivalent but of opposite sign.

14 We test whether wage innovations are serially correlated, and find some evidence that this is the case for highlyeducated women. We discuss the implications of this result in Section 3.5.


The second related model (Topel-IV, from now on) arises by noting that the error term in Eq. (1)contains an individual specific component in addition to the job specific component. The lattercomponent drives the bias detailed in (6) and (7), but this prior discussion omits the potential biasthat could result from heterogeneity in the individual specific component. Theoreticalconsiderations imply that individual heterogeneity will bias down the estimated return toexperience and bias up the estimated return to tenure.15 This is an important source of bias toaddress in our gender analysis since the extent of individual heterogeneity is likely to differ acrosswomen and men. Topel (1991) proposed instrumenting initial experience with actual experience toremove the bias from individual heterogeneity in the second step. Hence the Topel-IV procedure isdefined analogously to the Topel-2S estimator by Eqs. (4) and (5), as earlier, but X0 in Eq. (5) isinstrumented by X. We present estimates from this IV procedure to correct for potential bias thatcould stem from individual heterogeneity in addition to the estimates from Topel-2S.

The third and final model is based on Altonji and Shakotko (1987) (AS, from now on) whoproposed an instrumental variables estimator to address bias problems in estimating returns to tenuredue to individual and jobmatch heterogeneity in the wage equation. Themain instrumental variablesfor the tenure variables Tijt and Tijt

2 are the deviations of the tenure variables around their means forthe sample of observations on a given jobmatch. By construction, these instruments are uncorrelatedwith both the individual specific error component (μi) and the job match error component (ϕij). LetPTij be the mean of tenure for individual i over the sample observations in job j. Define T̃ ijt to be the

deviation of Tijt from the job mean, with T̃ ijt =Tijt−PTij . Let T̃ ijt

� �2¼ T2

ijt �PTij� �2

. The instrumental

variables procedure is a two stage least squares estimator in which T̃ ijt and (T̃ ijt)2 serve as instru-

mental variables for T and T2 in Eq. (1). If Tijt is also uncorrelated with the transitory errorcomponent ηijt then it is a valid instrumental variable. Due to the correlation between generalexperience and job match heterogeneity, the AS estimator is also biased.

The expression of the bias is:

bb2 � b2 ¼ �b1 � gXT b1 þ b2ð Þ ð8Þwhere b1 and b2 are as defined in Eq. (3) above.

Clearly neither theAS nor the Topel-2S estimators are ideal. The advantage ofAS is that it uses aninstrument that is exogenous to tenure. But it does not address the endogeneity of actual labor marketexperience. This procedure is also more susceptible to measurement error since the instruments aredeviations from the mean of job-specific tenure. Although Topel-2S estimates of returns to tenureand experience are biased, they are well defined in the sense that they are of the same magnitudealbeit of opposite sign. Moreover, in the absence of heterogeneity of within-job wage growth rates,the first step yields a consistent estimate of the sum of the returns to tenure and experience.

3.3. Gender disparities in labor market and job attachment

Before presenting our estimates of wage returns to tenure and experience, we first presentsome supporting evidence on gender disparities related to labor market and job attachment.

3.3.1. TurnoverSince our explanation of gender differences in wage returns to tenure and experience is based

on the hypothesis that women are less attached to their employers, we first present some direct

15 These opposing biases are discussed in both Topel (1991) and Altonji and Williams (2005).

Table 2Gross Annual Turnover Rates, NLSY 1979–1994 (Number of person-years in parentheses)

High School or Less More than High School

Separation Quit Laidd Off Separation Quit Laidd Off

Men 36% (10, 438) 23% (10, 438) 13% (4, 762) 31% (4,762) 23% (4, 4762) 8% (4,762)Women 40% (8,854) 30% (8,854) 10% (8,854) 35% (5,343) 28% (5,343) 7% (5,343)


evidence of gender disparities in quit and overall separation rates. Table 2 shows average turnoverrates for the NLSY sample that we use to estimate wage returns. The overall separation rates arehigher for women compared to men across both high and low education groups. These genderdisparities are even larger for quit rates. We confirm these patterns within a regression frameworkwhere we include a wide array of control variables. Table 3 presents the marginal effects from aProbit model specification. The coefficients on the female indicator variable show that quit ratesare higher for women than men in both education groups. Moreover, married women havesubstantially higher quit rates compared to their married male counterparts. The evidence onoverall separation rates confirms that married women have higher separation rates compared tomarried men and single women. However,for the high education group the coefficient on thefemale indicator variable is positive but insignificant. In the last two columns, we present thesame results for a subsample restricted to individuals who are at least 25 years of age. Theexclusion of very young workers ensures that the individuals in the more educated sample havecompleted their schooling and that women have entered their prime child-bearing years. Broadlyspeaking the coefficients show that both quit and separation rates are higher for women than menfor this subsample. Overall these results suggest that women are indeed less attached to theiremployers than men.

Studies by Becker and Lindsay (1994) and Sicherman (1996) confirm our finding that womenquit jobs at a higher rate than men. In addition, Light and Ureta (1995) find that women havelonger and more frequent non-working spells than men in their early careers, and that women tendto require relatively more time to accumulate a given amount of work experience. Banzhaf (2005)finds that women are less likely to remain in the labor market in the NLSY conditional on working

Table 3Turnover Regressions (marginal coefficients from Probit model specifications)

All Ages Age 25+ Years

Separation Quit Separation Quit

High School or LessFemale 0.006 (0.013) 0.080 (0.012) 0.011 (0.024) 0.075 (0.02)Married −0.053 (0.099) −0.040 (0.008) −0.078 (0.011) −0.047 (0.010)Female⁎Married 0.033 (0.014) 0.045 (0.013) 0.045 (0.017) 0.035 (0.015)Sample Size 22,896 22,896 13,145 13,145

More than High SchoolFemale 0.013 (0.018) 0.027 (.016) 0.082 (0.026) 0.061 (0.023)Married −0.057 (0.013) −0.022 (.012) −0.056 (0.013) −0.022 (0.012)Female⁎Married 0.064 (0.018) 0.043 (.016) 0.060 (0.019) 0.038 (0.017)Sample Size 14,710 14,710 10,723 10,723

Note: Controls variables include Age, Age Squared, Tenure, Experience, their interaction terms with Female, and Yearindicators. Standard errors are in parentheses below the coefficient estimates.

Table 4Training Incidence and Intensity: All Formal Training Programs versus Formal Company Provided Training Programs,NLSY 1986–1993

Obs All Training Company Provided Training

Incidence Total Hours Hours if Training Incidence Total Hours Hours if Training

High School or LessMen 336 26.5% 46.9 179 20.5% 25.4 125Women 303 27.4% 36.0 127 18.2% 13.9 69

More than High SchoolMen 313 47.6% 72.4 158 42.8% 59.4 149Women 298 40.3% 40.0 93 33.9% 24.7 75


in the previous period and less likely to enter the labor market if previously not employed. All ofthese findings support the view that women are relatively less attached to their jobs. To the extentthat women expect, and do experience, shorter job durations, the lower estimated rate of return ontenure for women is consistent with our strategic choice hypothesis.

In contrast, studies by Farber (1994) and Royalty (1998) of young women and men have notfound significant gender differences in separation probabilities. Note that both these studies focuson samples of workers who are younger than the workers in our sample. For example, Royalty(1998) includes in her sample individuals who may still be in school and where the oldestindividual is still less than 30 years of age. Imposing these same restrictions on the NLSY data wealso find that the coefficient on the female indicator variable is insignificant for more educatedworkers. Farber (1994) only includes individuals who have been in the labor market for at leastthree continuous years. This sample selection criterion clearly over samples women who are moreattached to the labor market. Our point of course is to understand whether women overall aremore or less attached to their jobs compared to men. Our analysis highlights significant genderdifferences in turnover for individuals in their middle twenties and thirties who are more likely tohave greater family commitments.

3.3.2. Incidence and intensity of trainingStarting in 1987, the NLSY contain detailed information on formal job training programs that

include the incidence and hours of training. More importantly, this training information isavailable for both company and non-company provided training programs. Since companyprovided training is more likely to be firm specific than non-company provided sources oftraining, we can ask whether women receive less company provided training than men.16

Table 4 shows the incidence and mean hours of total training and of company providedtraining. Since detailed training information is collected from 1987 onwards, our sample isrestricted to those who start a new job in 1987 (or after) and continue on the same job for twofurther periods. Hence the incidence rate represents the percent that receive any hours of trainingduring the first three survey periods on a specific job. Men receive more training than womenacross all subsamples. Average hours are higher for men than women for both total and company-

16 Our preliminary analysis is not a substitute for a rigorous analysis of gender disparities in training. Our objective is topresent summary statistics about gender disparities in company provided training that is consistent with the differentialwage returns to tenure across gender. Other studies (see below) have provided a more thorough analysis of trainingdisparities across gender.


provided training. In addition, men have a higher incidence of company provided training.Among more educated workers, men have higher incidence rate for training (48% versus 40%)and for company provided training (43% versus 34%) than women. The discrepancies are lesspronounced for the lower education group, although the training hours are greater for men thanwomen for all types of training. These differences across the high and low education groups areclearly consistent with the hypothesis that investments in training are greater among the moreeducated. This is also confirmed by the lower training incidence rate for the less educated groupoverall. The more dramatic finding is the difference in average hours of training across gender.Within the higher education group, men on average receive substantially more hours of totaltraining — over 50% more — than women. The gender disparity in total training among moreeducated workers is mainly due to the fact that men receive more than double the number ofaverage hours of company provided training than women—mean of 59 hours for men versus 25hours for women.

Royalty (1996) investigates whether gender disparities in company provided training could bedue to differences in expected job duration. She not only finds that predicted turnover issignificantly related to the likelihood of receiving company provided training, but that the positiveeffect of being male on the probability of receiving company training is reduced by about 25%when estimated turnover rates are included in the estimation. However, as also noted by Royalty,there remains some advantage to men in the receipt of company training that is not explained byobservables, including the turnover probabilities. Hence these training disparities could indeed bea consequence of employer discrimination.

The type and quantity of training women and men receive have important ramifications forreturns on tenure and experience. The observed gender differences in returns to tenure andexperience could indeed be the result of gender differences in training.17 There is an extensiveliterature on gender differences in training that suggests that women receive less training then men(Royalty, 1996; Olsen and Sexton, 1996; Hill, 1995; Lynch, 1992; Altonji and Spletzer, 1991).18

Further evidence from the NLSY is provided by Lynch (1992), who looks at a sample of non-college-graduates from 1980-83, and finds that males are more likely to receive on-the-jobtraining and to be in apprenticeships while women are more likely to receive off-the-job training.There are of course major limitations to using data from formal training programs, especiallygiven how little of such training is received. However, if reported training is just a small (butproportional) component of actual training then these reported differences in training may alsoreflect gender disparities in actual training levels. More importantly, note that our primary concernis on the more general issue of differential wage returns to tenure and experience since wagereturns reflect not only returns to formal training (as identified in the NLSY and in many of thetraining studies cited above), but also returns to informal training, learning by doing, and otherincentive-based compensation policies. As a consequence the major focus of our paper is on thegender differences in wage returns on tenure and experience.

17 Indeed, weaker job attachment implies that women are less likely to invest in on-the-job training. See Barron et al.(1993). Of course, gender differences in company provided training could be due to strategic investment decisions on thepart of the employee or due to employer discrimination.18 As mentioned earlier, gender disparities in wage returns on tenure could be due to factors other than just companyprovided training differentials such as back-loaded pay structures designed to elicit work effort or to incent self selectionof less mobile workers. Information on wage contracts and employment practices are, however, extremely rare, exceptperhaps in firm level data that includes personnel records.

Table 5Expectations of Labor Force Participation at Age 35, NLSY 1979–1984

Obs % Expecting to beWorking at age 35

% Not Expecting to be in Labor Forceat age 35 due to Family Reasons

High School or LessWomen 21,744 67.0% 26.9%Men 23,640 86.6% 1.7%

More than High SchoolWomen 13,734 79.9% 16.5%Men 12,060 91.8% 1.1%


3.3.3. Expectations about future labor force attachmentExpectations about future labor force participation will affect current human capital

investment decisions. Such expectations are likely to have implications for whether individualswill invest in general or specific forms of human capital. If an individual anticipates a longabsence from the labor market then she is more likely to make smaller overall investments.However, interruptions of a limited duration are unlikely to dampen overall levels of investment,but may have strategic impact on the type of investments. For example, a career interruption forfamily related reasons may imply a separation from an employer but not a substantial reduction inthe overall time spent in the labor market. Thus individuals who anticipate family related careerinterruptions in the future will be more likely to invest in general skills than in firm-specific skills.

A survey instrument in the NLSY allows us to study gender disparities in labor forceexpectations. From 1979 to 1984, respondents are asked, “What would you like to be doing at age35?” Table 5 presents the responses across women and men. First, in the second column wepresent the percent of women and men who expect to be working at age 35. Among the lesseducated workers 67% of women and 87% of men expect to be working. And among the moreeducated these numbers are 80% and 92%, respectively. Clearly, more educated women are a lotmore likely to report that they would be working at age 35 compared to less educated women(80% versus 67). This disparity across education is much smaller among men (92% versus 87%).The key point however is that these large disparities in labor force attachment are primarily due tothe fact that women expect to be out of the labor force for family related reasons. As the lastcolumn of Table 5 illustrates, less than 2% of men (in either education group) expect to be out ofthe labor force at age 35 due to family related reasons. By contrast, about 27% of less educatedwomen and 17% of more educated women expect to be out of the labor force due to family relatedreasons at age 35. These reports tell us that women indeed expect to have career interruptions intheir prime child-bearing and child-rearing ages, which in turn would imply that they are lesslikely to be attached to their employers.19

19 A striking illustration of rapidly changing work expectations of women is documented in two surveys of youngwomen aged 14 to 21 conducted a decade apart. In a 1969 survey less than 30% of the respondents expected to work atage 35, and just ten years later in 1979 over 70% of a similarly aged cohort of young women expected to work at age 35.No doubt, these dramatic changes in expectations are likely to have important implications for human capital investmentsof women (O'Neill, 1990).

Table 6Within-Job Wage Growth, NSLY 1979–1994 (Estimates from Topel’s First-Step)

Obs Estimated of β1+β2 Standard Errors

High School or LessWomen 4,792 0.079 0.0049Men 6,100 0.090 0.0049

More than High SchoolWomen 3,003 0.100 0.0074Men 2,894 0.118 0.0086


3.4. Wage returns to tenure and experience

Table 6 presents estimates of within-job wage growth from the first step of Topel's two-stepmodel. The model is under-identified by one parameter because first differences imply identicalunit changes for both experience and tenure. Therefore the constant term reflects the wageincrease due to an additional year of labor market experience — i.e. the joint wage returns totenure and general experience. Note that the overall return to labor market experience is higher formen than women in both the high and low education groups. An additional year of labor marketexperience is associated with about a 15 percent higher wage growth rate for men compared towomen in both education groups. This finding of lower overall returns to labor market experiencefor women has been documented in many previous studies (e.g., Loprest, 1992; Light and Ureta,1995). We now turn to our estimates of gender disparities in wage returns to tenure and experiencefrom the three models discussed in Section 3.2 earlier.

Table 7 presents the five-year cumulative returns to tenure and experience based on Topel-2S,Topel-IV, and AS procedures. These cumulative returns to tenure and experience, shown in thesecond and third columns, respectively, are presented for women and men across the high and loweducation groups, and they include the effects of higher order terms of tenure and experience.20

The main finding is that the returns to tenure are substantially smaller for women than they are formen, and that the returns to experience are somewhat larger for women than they are for men.These findings are robust across the different model specifications, and the gender differentialsare especially striking among more educated workers. For example, the estimates from Topel-IVshow that the five-year cumulative wage returns to tenure are almost three times greater for menthan they are for women, and that the five-year cumulative wage returns to experience are about40% more for women than they are for men. Note that the total return (shown in the last column)to an extra year of labor market experience is lower under AS than under Topel-2S and Topel-IVprocedures.21

A comparison of the estimates of gender disparities in returns to experience across Topel-2Sand Topel-IV among more educated workers (bottom panel of Table 7) show that the proposed

20 The full set of regression coefficients can be found in Appendix Tables 1 and 2.21 The total return under Topel-2S and Topel-IV in principle should be the same since they are based on Topel's first-step estimate. The slight disparities that appear in Table 7 are due to computations that approximate the cumulativereturns to tenure and experience, which are then simply added up to give the value presented in the last column entitledtotal return.

Table 7Five-Year Cumulative Wage Returns to Tenure and Experience

Obs Return to Tenure Return to Experience Total Returns

High School or LessTopel-2S Women 6,015 0.268 0.089 0.357

Men 6,873 0.291 0.116 0.407Topel-IV Women 6,015 0.167 0.183 0.350

Men 6,873 0.246 0.157 0.403AS Women 9,372 0.094 0.203 0.297

Men 11,245 0.099 0.181 0.280

More than High SchoolTopel-2S Women 3,367 0.228 0.261 0.489

Men 2,835 0.338 0.224 0.562Topel-IV Women 3,367 0.098 0.410 0.498

Men 2,835 0.272 0.288 0.560AS Women 5,645 0.072 0.255 0.327

Men 5,064 0.132 0.199 0.331

Note: These cumulative wage returns are calculated from regression coefficients of Topel-2S, Topel-IV, and ASprocedures, respectively. The detailed regression results are included in Table 1 and Table 2 of the Appendix.


heterogeneity correction in the latter procedure increases the return to experience for womensubstantially more (from .26 to .41) than it does for men (from .22 to .29). This finding impliesthat the extent of correlation between initial experience and individual heterogeneity is strongerfor women, and hence using actual experience as an instrument in Topel-IV increases the positivefemale-male gap in wage returns to experience. The corollary is that the negative gender gap inwage returns to tenure also increases since the return to tenure is estimated by subtracting thereturn to experience from the total return to an extra year of labor market experience. Thus thiscorrection for individual heterogeneity lends even more support to our hypothesis that womeninvest less in firm-specific skills and more in portable general skills, especially among moreeducated workers where strategic human capital investments are likely to matter more.

The key inference question is whether the observed gender differences in these wage returnsare significant. The computation of correct standard errors requires not only that the error terms inthe wage equation for men and women are uncorrelated, but also other stringent assumptions. Inparticular, we need to assume that the error terms from Topel's first and second steps areuncorrelated, and more significantly, that the estimates of the returns to tenure and experience areunbiased. Under these assumptions we compute standard errors that show that the marginalreturns to tenure and experience are significantly different across women and men. Thisconclusion holds for values of experience and tenure up to five years for the Topel-IV estimates.The problem of course with these computed standard errors is that we know from theoreticalconsiderations that the estimates of returns to experience and tenure are biased. Since biasedestimates are composed of the true estimate and a bias term, computing an appropriate standarderror requires that we make arbitrary assumptions about the variance of the bias term. Theseconsiderations temper any pretense of providing a conclusive statistical test of gender differencesin the point estimates of the returns to tenure and experience. However, the discussion in the nextsection on whether the systematic biases inherent in our estimates are different across women andmen suggests that our estimated gender disparities in wage returns to tenure and experience are infact a lower bound of the true gender disparities.


3.5. Potential biases

In this section we consider some potential biases of our estimates of returns to tenure andexperience that are likely to differ systematically across women and men, and assess theirimplications for our estimates of the gender disparities in returns to tenure and experience. Inparticular, we focus on biases that could arise from gender differences in job search effort andheterogeneity of within-job wage growth rates, and present evidence to show that the true gendergap in returns to tenure are likely to be even larger than what our estimates suggest.22

3.5.1. Gender differences in job searchIn Section 3.2 we claimed that the estimate of returns to experience is biased due to job search.

The standard assumption in job matching and job search theories is that workers sample new joboffers from a stable distribution and that offers arrive randomly at an exogenous rate. If the truereturns to experience and tenure are zero — β1=0 and β2=0 — then this optimal job searchprocess implies that the current wage of an individual is the maximum of all the job offersreceived. Since the number of job offers increase with experience, current wages are correlatedwith experience even though we have assumed that the true return to experience is zero. This jobsearch process is the reason that we suppose that our return to experience is biased — i.e. b1N0.Note that this argument (based on Topel, 1991) is predicated on the simplifying assumption thatoutside job offers arrive at an exogenously given rate. Workers however are likely to adjust theirsearch effort in order to elicit job offers at different rates, and it is likely that search effort for newjobs will be systematically different across men and women for precisely the same reasons thatwomen are less likely to invest in job specific skills. If women are less attached to jobs then theywould invest less in job search compared to their male counterparts, implying a weakercorrelation between initial experience and the bias term on returns to experience (b1).

23 Thus thebias in our estimates of returns to experience would be smaller for women than men. The im-plications of endogenous job search effort for differences in wage returns between women andmen are obvious: our positive estimate of the female-male gap in returns to experience is likely tobe bigger, and correspondingly, the large negative gap in returns to tenure is likely to be anunderestimate of the true gap. Moreover, if women value other non-pecuniary dimensions of jobsmore than men then even the same levels of search effort are unlikely to generate comparablewage offers across women and men, implying again, a weaker correlation between experience andthe bias term on returns to experience (b1).

One implication of lower levels of job search effort and placing relatively more weight on jobcharacteristics such as hours and location of work than wages (Kahn and Griesinger, 1989) is thatmobility wage gains should be smaller for women than men. Loprest (1992) using the sameNLSY data shows that mobility wage gains for women are about half of what they are for men,and attributes part of the gender discrepancy to choice of hours, location and occupation. Lightand Ureta (1995) also provide evidence that mobility wage gains are smaller for women. Inaddition, we confirm that this pattern holds for our sample, especially for workers between the

22 Bratsberg and Terrell (1998) have a more extensive discussion on the sensitivity of both the AS and Topel estimatorsto various considerations that could lead to bias in their estimators, including choice of metric of tenure and experience.Using the same NLSY data, they argue that their key conclusion — black workers receive a comparable or even higherreturn to tenure than white workers, but earn a far lower return to general labor market experience— holds because thesebiases do not necessarily compromise the comparison across races. Our discussion here focuses on considerations wethink might lead to biases that differ systematically across women and men.23 Sandell (1980) presents some evidence to suggest that unemployed married women invest too little in job search.


ages of 20 and 30. This gender difference in mobility wage gains only strengthens our findings ongender gaps in returns to tenure and experience since it suggests that the true gender disparities inreturns to experience and tenure are even larger than the gaps we have estimated.

3.5.2. Heterogeneity of within-job wage growthIn Section 3.2 we claimed that Topel's first step generated consistent estimates of overall

returns to an extra year of labor market experience if within-job wage increases are seriallyuncorrelated. In this section we address whether within-job wage increases are indeed seriallyuncorrelated for women and men. For example, if jobs differ in wage growth rates and the sourceof wage growth is due to firm-specific factors then high growth jobs are more likely to survive. Asa consequence, Topel's first step would over estimate the return to an extra year of labor marketexperience because our data would over represent high growth jobs. The key implication ofheterogeneity of within-job wage growth rates is that it would tend to bias up our estimate ofreturns to tenure. Note that we obtain the return to tenure by subtracting the estimate of returns toexperience from our estimate of the overall within-job wage increase from one period to the next.Our objective in this section is to test whether there is evidence of such heterogeneity, and if so,whether the bias resulting from it might be different across women and men.

To test for within-job wage growth heterogeneity we follow a test proposed by Topel. Hisexposition begins by rewriting the wage growth model in Eq. (4):

yijt � yijt�1 ¼ b1 þ b2 þ gijt þ mijt � mijt�1;

where ηijt=/ijt− /ijt−1. Note that if this permanent component of wage increase is specific to anemployment relationship then high wage growth jobs are more likely to survive. In the observedsample of within-job wage growth we would expect Eη N 0 if high (low) anticipated values of ηijtare more (less) likely to survive in the future. Denote Rt as the remaining life a job from time t.Then the following condition

0 V E gtjRt z 0ð Þ b E gtjRt z 1ð ÞbE gtjRt z 2ð Þ b N ;

says that the expected value of wage innovations are larger for jobs that survive for longer timeperiods. A simple procedure to test whether wage innovations are larger for longer duration jobs isto assess whether the residual of current wage growth is positively correlated with the remaininglife of the job as indicated in the above expression (Topel, 1991). To implement this procedure wemodify Topel's first step in Eq. (4) by first adding a linear term to indicate the remaining time lefton the job. To allow for non-linear effects we can of course include indicator variables to estimateseparate effects of jobs that end in period t+1, t+2, t+3, and t+4. Table 8 presents evidence ofwithin-job wage growth rates across jobs that end in the near and distant future. The first columnlabeled “t” tests whether the relationship between remaining tenure and current wage growth islinear. The remaining 4 columns are the coefficients on indicator variables for remaining years onthe current job. Since jobs that last for more than 4 periods are the omitted category, negative anddecreasing coefficients on these time indicator variables would be evidence of within-job wagegrowth heterogeneity. Except for the group of women with more education, we find no evidenceof the above inequalities.24 This suggests the presence of heterogeneity of wage innovations for

24 We find, like Topel and Ward (1992), that the evolution of within-job wages follows a random walk for men.Moreover, we also find that there is no evidence of positive serial correlation of within-job wage innovations for less-educated women.

Table 8Evidence on Heterogeneous Within-Job Wage Growth Rates (Relationship between Remaining Job Duration and CurrentWage Growth)

Remaining Job Duration in Years

t t+1 t+2 t+3 t+4

High School or LessWomen (Obs=4,759) 0.001 (0.001) −0.005 (0.007) −0.005 (0.008) 0.006 (0.009) 0.003 (0.01)Men (Obs=6,048) 0.001 (0.001) −0.006 (0.006) −0.018 (0.007) −0.0003 (0.008) −0.010 (0.009)

More than High SchoolWomen (Obs=2,985) 0.004 (0.001) −0.027 (0.009) −0.035 (0.010) −0.024 (0.012) −0.021 (0.013)Men (Obs=2,869) 0.0008 (0.0015) −0.003 (0.009) −0.006 (0.010) 0.0005 (0.012) −0.009 (0.013)

Note: Standard errors are in parentheses.


our sample of more educated women. The preceding discussion then implies that our estimate ofwage returns to tenure for more educated women is likely to be an over estimate, suggesting thatthe implied gender gap in wage returns to tenure is likely to be even larger than the gaps weestimate in Section 3.4.

Both types of bias we have considered in this section suggest that we are overestimating thereturns to tenure for women compared to men. This leads to the conclusion that especially forhighly educated workers, the gender difference in the return to tenure is even larger than what ourestimates suggest.

These findings must be tempered with the possibility that under certain wage settingconditions lack of positive serial correlation in within-job wage growth is an inappropriate testto reject the hypothesis of heterogeneous wage growth rates among jobs (Munasinghe, 2006).Thus Topel's first step may be an over estimate of the overall returns to an additional year oflabor market experience even if we have no evidence of positive serial correlation of within-jobwage increases. However, these wage setting considerations also that show that the tenure-wage profile will be flatter than the tenure-productivity profile. Thus the over sampling ofhigher wage growth jobs is ameliorated by the fact that any estimate of the tenure effect onwages is likely to be an underestimate of the underlying increase in firm-specific productivity.The key point for our purposes is that the gender disparity in wage returns to tenure isnevertheless informative of differences in firm-specific skill accumulation across women andmen.

4. Conclusion

We estimate returns to experience and tenure across a sample of women and men in their earlycareers. We confirm that the overall wage return to an additional year of labor market experienceis higher for men than women. However, a decomposition of this overall return shows thatespecially among more educated workers, the return on tenure is substantially lower for womenthan for men, but that the return on experience is higher for women than men. These findings areconsistent with the hypothesis that women in comparison to men might be less attached to jobsdue to life cycle events such as marriage and child birth. Our interpretation is also compatible withsome of the findings in Light and Ureta (1995), namely that for women the drop in earnings uponreturning to work after a career interruption is smaller than it is for men, and the “catch up” to their


continuously employed counterparts is quicker. Further we show that women are more likely toquit their jobs than men, that the gender disparity in job training intensity is due to the fact thatmen receive substantially more company provided training than women, and that a much higherfraction of women expect to be out of the labor force at age 35 due to family related reasons. Thisbody of evidence points to the fact that women are indeed less attached to their employers thanmen.

This paper suggests various further research directions. The first is to extend our empiricalanalysis by explicitly introducing instruments for women's commitment to the labor market andto specific jobs. For example, worker's expectations about job duration, future labor forceparticipation, care for parents, and the timing of marriage and children can be utilized to refine theanalysis presented in this paper. How do investments and the returns to experience differ acrosswomen who anticipate working in the labor market versus those who do not? Do women whoexpect to marry and have children later in life experience greater returns to labor marketexperience? Do women who expect long job durations experience higher returns to tenure?Answers to these questions could potentially help to clarify further whether the observed genderdisparities in the wage returns to experience and tenure are a consequence of investment choice ordiscrimination on the part of employers.

Appendix

Table 1Log Wage Regressions (High School or Less), NLSY 1979–94

Variable Topel-2S AS Topel IV

Women Men Women Men Women Men

Tenure 0.0621(0.0051)^

0.0649(0.0051)^

0.033(0.006)

0.0316(0.005)

0.0456(0.0055)^

0.0578(0.0053)^

Tenure2 −0.0029(0.0005)

−0.0028(0.0004)

−0.003(0.0004)

−0.003(0.0004)

−0.0029(0.0005)

−0.0028(0.0004)

Experience 0.0169(0.0015)

0.0254(0.0014)

0.0385(0.004)

0.036(0.004)

0.0335(0.0024)

0.0326(0.002)

Experience2 0.00002(0.0004)

−0.0007(0.0004)

−0.0003(0.0003)

−0.0005(0.0002)

0.00002(0.0004)

−0.0007(0.0004)

Union 0.099(0.014)

0.274(0.011)

0.144(0.011)

0.275(0.008)

0.112(0.014)

0.275(0.011)

SMSA 0.166(0.010)

0.103(0.009)

0.132(0.008)

0.093(0.008)

0.160(0.010)

0.101(0.009)

Health −0.067(0.026)

−0.048(0.035)

−0.074(0.022)

−0.0265(0.026)

−0.066(0.026)

−0.046(0.035)

Hrs worked/wk 0.0046(0.026)

0.002(0.0005)

0.004(0.0004)

0.0018(0.0004)

0.004(0.0005)

0.0019(0.0005)

AFQT 0.0033(0.0002)

0.0023(0.0002)

0.0033(0.0002)

0.0027(0.0001)

0.0033(0.0002)

0.0024(0.0002)

Constant 1.125(0.026)

1.455(0.021)

1.122(0.022)

1.464(0.026)

1.051(0.027)

1.355

Obs 6,015 6,873 9,372 11,245 6,015 6,873Adjusted R2 0.1679 0.2203 0.2689 0.3068 0.151 0.2170

^ Estimated upper bound for standard errors.

Table 2Log Wage Regressions (More than High School), NLSY 1979–94

Variable Topel-2S AS Topel IV

Women Men Women Men Women Men

Tenure 0.0468(0.0078)^

0.0692(0.009)^

0.0281(0.0079)

0.0371(0.0087)

0.0244(0.008)^

0.0590(0.009)^

Tenure2 −0.0012(0.0006)

−0.0022(0.0007)

−0.0028(0.0006)

−0.0024(0.0008)

−0.0012(0.0006)

−0.0022(0.0007)

Experience 0.0533(0.00246)

0.0489(0.0026)

0.0453(0.0065)

0.0383(0.007)

0.0757(0.004)

0.0591(0.004)

Experience2 −0.0014(0.0005)

−0.0017(0.0005)

0.00001(0.0004)

−0.0004(0.0004)

−0.0014(0.0005)

−0.0017(0.0005)

Union 0.132(0.031)

0.127(0.027)

0.1602(0.0203)

0.1295(0.0187)

0.136(0.031)

0.126(0.027)

SMSA 0.209(0.0184)

0.195(0.021)

0.165(0.0136)

0.1959(0.0147)

0.205(0.019)

0.191(0.021)

Health −0.064(0.050)

−0.174(0.072)

−0.103(0.0353)

−0.178(0.0521)

−0.045(0.050)

−0.166(0.073)

Hrs worked/wk 0.003(0.0007)

−0.002(0.0008)

0.0024(0.0006)

−0.0018(0.0006)

0.0027(0.0008)

−0.0023(0.0008)

AFQT 0.002(0.0003)

0.0014(0.0004)

0.0034(0.0003)

0.002(0.0003)

0.0021(0.0004)

0.0014(0.0004)

Constant 1.244(0.051)

1.751(0.057)

1.294(0.041)

1.755(0.0461)

1.124(0.052)

1.683(0.059)

Obs 3,367 2,835 5,645 5,064 3,367 2,835Adjusted R2 0.2373 0.2678 0.2742 0.2929 0.2192 0.02639

^ Estimated upper bound for standard errors.


References

Altonji, Joseph, Shakotko, Robert, 1987. Do Wages Rise with Job Seniority. Review of Economic Studies 54, 437–459(July).

Altonji, Joseph, Spletzer, James, 1991. Worker Characteristics, Job Characteristics and the Receipt of On-the-Job Training.Industrial and Labor Relations Review 45 (1), 58–79.

Altonji, Joseph, Blank, Rebecca, 1999. Race and Gender in the Labor Market Chapter 48. In: Ashenfelter, O., Card, D. (Eds.),Handbook of Labor Economics, vol. 3. Elsevier Science.

Altonji, Joseph, Williams, Nicolas, 2005. Do Wages Rise With Job Seniority? A Reassessment. Industrial and LaborRelations Review 58 (3), 370–397.

Banzhaf, Melissa Ruby (2005), Structural Differences in Job Turnover and Wage Growth by Sex and Education, Ph.D.Dissertation. Chapel Hill, NC: University of North Carolina.

Barron, John, Black, Dan, Loewenstein, Mark, 1993. Gender Differences in Training, Capital and Wages. Journal ofHuman Resources 28 (2), 343–364.

Becker, Elizabeth, Lindsay, Cotton, 1994. Sex Differences in Tenure Profiles: Effects of Shared Firm-SpecificInvestments. Journal of Labor Economics 12 (1), 98–118.

Blau, Francine, Kahn, Lawrence, 1997. Swimming Upstream: Trends in the Gender Wage Differential in the 1980s.Journal of Labor Economics 15 (1, part 1), 1–42.

Bratsberg, Bernt, Terrell, Dek, 1998. Experience, Tenure, and Wage Growth of Young Black and White Men. Journal ofHuman Resources 33, 658–682.

Farber, Henry S., 1994. The Analysis of Interfirm Mobility. Journal of Labor Economics 12, 554–593 (October).Filer, Randall, 1993. The Usefulness of Predicted Values for Prior Work Experience in Analyzing Labor Market Outcomes

for Women. Journal of Human Resources 28 (3), 658–682.Hashimoto, Masanori, 1981. Firm-Specific Human Capital as a Shared Investment. American Economic Review 71 (3, June),

475–482.


Hersch, Joni, Reagan, Abagail, 1997. Worker Effort Decisions and Efficient Gender-Specific Wage-Tenure Profiles.Economic Inquiry 35 (1), 193–207.

Hill, Elizabeth, 1995. Labor Market Effects of Women's Post-School-Age Training. Industrial and Labor RelationsReview 49 (1), 138–149.

Kahn, Shulamit, Griesinger, Harriet, 1989. Female Mobility and the Returns to Seniority: Should EEO Policy beConcerned with Promotion? Papers and Proceedings American Economic Review 79 (2), 300–304 May.

Kim, Moon Kak, Polachek, Solomon, 1994. Panel Estimates of Male-Female Earnings Functions. Journal of HumanResources 29 (2), 406–428.

Lazear, Edward P., 1981. Agency, Earnings Profile, Productivity, and Hours Restrictions. American Economic Review 71,606–620 (September).

Light, Audrey, Ureta, Manuelita, 1990. Gender Differences in Wages and Job Turnover Among Continuously EmployedWorkers. Papers and Proceedings American Economic Review 80 (2), 293–297 May.

Light, Audrey, Ureta, Manuelita, 1995. Early-Career Work Experience and Gender Wage Differentials. Journal of LaborEconomics 13 (1, January), 1995.

Loprest, Pamela, 1992. Gender Differences in Wage Growth and Job Mobility. American Economic Review 82 (May).Lynch, Lisa, 1992. Private Sector Training and the Earnings of Young Workers. American Economic Review 82, 299–312

(March).Munasinghe, Lalith, 2006. A Theory of Wage and Turnover Dynamics. Working Paper. Barnard College, Department of

Economics, New York.Olsen, Reed Neil, Sexton, Edwin, 1996. Gender Differences in the Returns to and the Acquisition of on-the-job Training.

Industrial Relations 35 (1), 59–77.O'Neill, June, 1990. Women and Wages. The American Enterprise 1.6 (Nov.–Dec.), 25–33.O'Neill, June, Polachek, Solomon, 1993. Why the Gender Gap in Wages Narrowed in the 1980's. Journal of Labor

Economics 11 (Part I, January), 205–228.Royalty, Anne Beeson, 1996. The Effects of Job Turnover on the Training of Men and Women. Industrial and Labor

Relations Review 49 (3), 506–521.Royalty, Anne Beeson, 1998. Job-to-Job and Job-to-Nonemployment Turnover by Gender and Education Level. Journal of

Labor Economics 16 (2), 392–443.Salop, Joanne, Salop, Steven, 1976. Self-Selection and Turnover in the Labor Market. Quarterly Journal of Economics 90

(November).Sandell, Steven H., 1980. Is the Unemployment Rate of Women Too Low? A Direct Test of the Economic Theory of Job

Search. Review of Economics and Statistics 62 (4, November), 634–638.Sicherman, Nachum, 1996. Gender Differences in Departure from a Large Firm. Industrial and Labor Relations Review 49

(3, April), 1–2.Spivey, Christy, 2005. Time Off at What Price? The Effects of Career Interruptions on Earnings,q. Industrial and Labor

Relations Review 59 (1), 119–140.Topel, Robert, 1991. Specific Capital,Mobility, andWages:WagesRisewith Job Seniority. Journal of Political Economy 9 (1),

145–176.Topel, Robert, Ward, Michael P., 1992. Job Mobility and the Careers of Young Men. Quarterly Journal of Economics

145–176 (May).Waldfogel, Jane, 1998a. The Family Gap for YoungWomen in the United States and Britain: Can Maternity Leave Make a

Difference? Journal of Labor Economics 16 (3), 505–545.Waldfogel, Jane, 1998b. Understanding the ‘Family Gap’ in Pay for Women with Children. Journal of Economic Perspectives

12 (1), 137–156.Wellington, Alison, 1993. Changes in the Male Female Wage Gap (1976–95). Journal of Human Resources 28 (2),

383–411.

gender gap in wage returns to job tenure and experience · of gender differences in wage returns to...

Documents